Training Deep Learning Segmentation Models Using Super-Resolution Crack Images for Detection of Thin Concrete Cracks

Cracks are the most critical damage types on concrete structures, and they are commonly assessed based on visual inspection. Recently, there have been many attempts to replace conventional inspection with computer vision-based inspection powered by deep learning. However, the widths of cracks are very thin in the order of submillimeters, which requires high-resolution (HR) images for inspection. However, the low resolution of crack images due to the limitations of imaging devices and poor accessibility to structural members is one of the biggest hurdles to accurately detecting thin cracks in the field. This study proposes a training framework for deep learning segmentation models to improve the performance of crack segmentation in low resolution (LR) images by employing a super-resolution (SR) model. Different from the conventional use of SR models in the testing images, the proposed framework applies the SR model in both training and testing to minimize the false positives and negatives generated by the SR model. In this study, two single-image super-resolution (SISR) models (i.e., enhanced deep residual (EDSR) and Super-resolution generative adversarial network (SRGAN)) and two famous semantic segmentation models (i.e., CGNet and DeepLabV3+) were used to validate the performance of the proposed framework. Thirteen experimental scenarios were designed using different training and testing data combinations of the HR, LR, and SR images obtained from the concrete pavement. The crack detection performance of each scenario was evaluated using the intersection over union (IoU) and its ratio over the IoU of the best scenario employing HR images in both training and testing. The performance of the proposed framework using SR images in both training and testing ranged between 85% and 90% compared to that of the topline scenario using HR images in both training and testing, which shows high applicability to the detection of thin cracks in LR images commonly obtained from real civil structures.